Ultravox

Ultravox is a real-time voice AI infrastructure layer that provides developer-friendly APIs and SDKs to build natural, low-latency voice agents. By using a speech-native model rather than traditional text-to-speech pipelines, it preserves paralinguistic cues like tone and cadence while minimizing conversation delays. The platform offers telephony support, built-in scaling tools, and flexible pricing tiers ranging from pay-as-you-go to enterprise plans.

• Developer-friendly REST APIs for easy integration
• Intuitive developer kits and SDKs for every major platform across web and mobile
• Built-in integrations with the largest telephony providers
• Dynamic endpointing through a neural VAD model that predicts conversation states and turn-taking
• Open weight models available on Hugging Face
• Outbound Call Scheduler
• Custom voice clones
• Corpora for RAG

Pros

• Speech-native modeling preserves paralinguistic signals like tone, pitch, and cadence without relying on traditional speech-to-text transcription
• Managed full inference stack and purpose-built infrastructure minimize conversation latency
• Flexible tiering accommodates experimentation with a free starting option as well as dedicated scaling and enterprise plans

Cons

• The Pay As You Go plan includes hard limits on concurrency, restricting users to up to 5 concurrent calls
• Certain features like custom voice clones, corpora for RAG, and the outbound call scheduler vary by plan tier
• Enterprise support and priority SLAs require a custom-priced enterprise plan

Ultravox fits developers and engineering teams building real-time, conversational voice agents for applications requiring low latency and natural turn-taking. It is well-suited for organizations scaling voice AI systems across web, mobile, and telephony channels. It is a poor fit for teams looking for traditional text-to-speech pipelines that rely on intermediate text transcription or projects needing zero-infrastructure management.

Ultravox operates on a freemium model with a free tier and paid plans above it. The Pay As You Go tier costs $0/month, includes 30 minutes of free calls, allows up to 5 concurrent calls, and charges $0.05 per minute after the initial free minutes, with threads priced at $2.00 per 1M uncached input tokens and $15.00 per 1M output tokens, and no charge for cached input tokens. The Pro plan costs $100/month when billed yearly, removes hard caps on concurrency, and adds features such as an outbound call scheduler, 5 custom voices, and 20 corpora for RAG at a rate of $0.05 per minute. The Enterprise plan is custom-priced and designed for massive scale, featuring priority SLAs, org support, and customizable options.

Unlike traditional voice AI systems that convert speech to text before an LLM can process it, Ultravox uses a speech-native model trained from first principles. By avoiding intermediate text transcription, the platform preserves critical paralinguistic signals such as tone, cadence, and pitch while eliminating the translation latency that makes standard voice agents feel slow and robotic. Additionally, Ultravox manages its own end-to-end inference stack and purpose-built infrastructure rather than relying on external LLMs or shared inference pools.

Ultravox works with major platforms across web and mobile via SDKs, provides REST APIs, and includes built-in integrations with the largest telephony providers.